Claude-Flow / Ruflo: Year One Executive Brief — 31 verified first-mover positions, harness-not-model thesis, top 20 innovations with calibrated first-claim status Claude-Flow, an orchestration harness built by developer Reuven Cohen around Anthropic's Claude Code and OpenAI's Codex, has established 31 verifiable first-mover positions in its first year. The project's thesis holds that the future lies in the harness—coordination, memory, verification, governance, routing, and federation—rather than the model itself. None of the 31 firsts involve model architecture, reinforcing the strategic focus on substrate-agnostic scaffolding. 31 verified firsts. One thesis. June 2025 – June 2026 Author: ruvnet Reuven Cohen · Compiled 2026-06-13 claude-flow is the orchestration harness around Anthropic's Claude Code and OpenAI's Codex — distributed as npx claude-flow , npx @claude-flow/cli , and the user-facing wrapper npx ruflo . It is what wraps the model, schedules the agents, persists the memory, signs the evidence, enforces the budget, and routes the work. In its first twelve months — between the initial commit on 2025-06-02 and the v3.10.46 stable release on 2026-06-13 — the project established 31 verifiable first-mover positions in the agent-orchestration category. Five of these are clean prior-art claims with no comparable public artifact found at compile time witness manifests with temporal verification, cross-installation agent federation with trust scoring, signed browser-session cognitive artifacts, hierarchical-mesh topology with distributed-systems consensus in an agent CLI, and routed deterministic Tier-1 execution before the LLM call . Every claim is anchored to a public commit URL and an npm publish timestamp. The project shipped, in production-tagged npm releases: 11,591 commits across one canonical repo 314 MCP tools exposed to Claude Code and Codex 16 agent roles + custom types coordinator, architect, researcher, coder, tester, reviewer, security-architect, memory-specialist, performance-engineer, and 7 more 19 AgentDB controllers spanning routing, recall, causal graphs, context synthesis, and pattern store 21 native plugins in the published ruflo distribution, covering ADRs, browser harness, federation, IoT devices, knowledge graphs, neural trading, and more 3 npm packages version-aligned across latest , alpha , and v3alpha dist-tags The thesis driving these twelve months — and the reason the work matters — is this: The future is not the model. The future is the harness.Nearly every claude-flow capability is about coordination, memory, verification, governance, routing, provenance, or federation — not model architecture. The model is a substrate; the harness is the product. Year one shipped the substrate-agnostic harness. The 31 verified firsts in the source dossier sort cleanly by what they actually do . The split is striking: | Theme | Firsts | Examples | |---|---|---| Coordination | 6 | Swarm topology hierarchical/mesh/adaptive , Hive-Mind queen-led, Raft/BFT/Gossip/CRDT consensus, Maestro, dual-mode Claude+Codex, Agent Teams + SendMessage | Memory | 6 | AgentDB + HNSW, RaBitQ 1-bit quantization, Claude Code memory bridge ADR-076 , RVF cognitive containers, ReasoningBank trajectory store, transfer-hook IPFS pattern sharing | Verification & Provenance | 5 | Witness manifests with temporal history ADR-103 , signed browser trajectories ADR-122 , CI-enforced witness verification, verifiable corpus benchmarks, AIDefence with OWASP ASI01 mapping ADR-131/132 | Governance & Routing | 5 | 3-tier model routing ADR-026 , deterministic Tier-1 codemods ADR-143 , PreToolUse modification hooks, ToolOutputGuardrail, federation-wide budget circuit breaker ADR-097 | Federation | 3 | Agent LLM Federation plugin ADR-078 , IPFS plugin registry, IPFS cross-project pattern sharing ADR-021 | Cross-domain primitives | 4 | IoT Cognitum device-as-swarm-member, agentic-payments BFT, neural-trader, ruflo-goals horizon tracking | Methodology & UX | 2 | SPARC 5-phase automation, statusline observability ADR-052 | Model architecture | 0 | — none. claude-flow does not ship a model. | Zero of the 31 firsts are about the model. Every one of them is harness work — the scaffolding that turns a stochastic LLM call into something that can be coordinated, remembered, verified, governed, routed, and federated. This is the strategic posture: the harness is the durable surface, the model is interchangeable substrate. This matters commercially because model capability is now a moving target with quarterly resets. The harness is what accumulates value across model generations. A coordination pattern shipped in June 2025 still works against a Claude Opus released in June 2026. A witness manifest signed yesterday is still verifiable tomorrow. The harness is where compound interest happens. The twenty most strategically significant firsts from the source dossier. The opening five are the strongest verifiable prior-art positions — the ones that anchor the harness thesis with cryptographic, content-addressed, or topology-level evidence that competing harnesses do not have. Each entry carries a First-claim status line so the position is legible at a glance. The first-claim status uses three forms, deliberately calibrated: Verified first — no prior public artifact found through commit-log search, npm registry inspection, and external-coverage review at compile time Among the earliest — predates one or more comparable systems and shipped concurrent with or before the broader category wave First in the Claude-Code harness category — first within the specific surface of orchestrating Claude Code and later Codex as the host harness Every claim links to the public commit. Anyone can re-verify. A cryptographically signed Ed25519 fix manifest that records, with a timestamp and a content hash, that a specific fix is present in the working tree — and gives CI a one-command way to verify it across releases. If a fix regresses, the manifest diff identifies the introducing commit. First commit: — 2026-05-09 76bd35943 CI integration: — 2026-05-09 54c706f56 npm: claude-flow@3.6.x May 2026 First-claim status: Verified first in the agent-harness category. No prior public artifact found via commit-log search across comparable harnesses AutoGen, CrewAI, LangGraph, OpenAI Swarm, HAL at compile time. Signed temporal-history fix manifests with CI-enforced verification are a category of their own. Strategic significance: Signed-provenance for "this fix actually shipped" is more rigorous than any CHANGELOG entry. It converts release notes from prose into cryptographic claim. Cross-installation agent federation with zero-trust security, trust-score-gated message routing, and per-peer budget circuit breakers. Independent claude-flow installs can collaborate without a central server, and the federation enforces spend limits per peer. First commit: — 2026-04-29 f3ea2d778 Trust scoring: — 2026-04-29 41957d0f1 Budget circuit breaker ADR-097 : — 2026-05-04 62a6fc5fb npm: claude-flow@3.6.x / ruflo@3.6.x April 2026 First-claim status: Verified first for cross-installation agent federation with trust-score-gated routing and per-peer economic circuit breakers. No comparable public artifact found in commit-log search or external coverage at compile time — the closest adjacent work multi-tenant SaaS agent platforms is a different architecture entirely. Strategic significance: Federation between independent installs — with cryptographic peer auth and economic constraints — is a different model from multi-tenant SaaS. It is the first agent-network primitive that does not require a central operator. Browser sessions captured into Ruflo Vector Format RVF containers with signed trajectories. A replay is verifiable; a session is a portable, content-addressed asset, not an ephemeral run. Browser plugin first commit: — 2026-04-28 81418649c Session-as-skill architecture: — 2026-05-04 b5b6fb3fb Signed trajectories ADR-122 : — 2026-05-18 5fedd1d02 npm: claude-flow@3.6.x / ruflo@3.6.x April 2026 First-claim status: Verified first for signed, content-addressed browser-session cognitive artifacts. Browser automation harnesses Playwright, Puppeteer, Browserbase capture sessions as ephemeral runs or video; capturing them into a portable, cryptographically signed, replayable vector container is, to public knowledge at compile time, unique. Strategic significance: Browser-automation agents are common; capturing the session into a signed, replayable, content-addressed format is rare. Every browse becomes evidence. Explicit --topology hierarchical|mesh|adaptive selection at the CLI, with hierarchical-mesh as the anti-drift default and pluggable consensus Raft, Byzantine, Gossip, CRDT, Quorum . First swarm subsystem: — 2025-06-11 c9b73c4f2 Consensus suite: — 2025-07-25 541fc88c3 Hierarchical-mesh hybrid: — 2026-01-04 1b6a1971f npm: claude-flow@1.0.22 2025-06-11 for swarm primitive; @3.0.0-alpha.1 2026-01-06 for hierarchical-mesh default First-claim status: Verified first in the Claude-Code harness category for the swarm primitive June 2025 predates the broader MCP-aware orchestrator wave ; verified first for pluggable distributed-systems consensus Raft/BFT/Gossip/CRDT exposed as agent-CLI flags in July 2025; verified first for the hierarchical-mesh hybrid topology as an anti-drift default January 2026 . Strategic significance: Distributed-systems consensus Raft, BFT inside an agent CLI is unusual. Topology as a first-class CLI flag, with hierarchical-mesh as the anti-drift default, is the coordination substrate everything else assumes. Replaces Agent Booster for intent-known transforms var-to-const , remove-console , add-logging with a TypeScript-compiler-driven deterministic path. Cost: $0. Latency: ~1ms. Result: verifiable, idempotent, LLM-free. First commit: — 2026-05-29 0988d92ce Wired into cost-trend: — 2026-05-29 0403a3427 npm: claude-flow@3.10.x late May 2026 First-claim status: Verified first in the Claude-Code harness category for a routed deterministic-execution tier TypeScript-compiler-driven, $0, ~1ms that sits before the LLM call rather than as a post-LLM optimization. Adjacent IDE tooling has codemod libraries; embedding them as Tier-1 routing in an agent harness with measured benchmark evidence is, at compile time, unique to this project. Strategic significance: Quantifies "skip the LLM where the transform has zero ambiguity" with a measurable corpus benchmark. The most expensive call is the one you never made. A hooks vocabulary covering pre/post-edit, pre/post-task, pre/post-command, session-start/end/restore, and intelligence hooks route, explain, transfer . Shipped before Anthropic formalized the hooks contract in Claude Code 1.0.51. First commit: — 2025-07-08 c8abe7459 Hook safety system: — 2025-07-08 93b0d31e3 npm: claude-flow@2.0.0-alpha.16 2025-07-08 First-claim status: Among the earliest — shipped July 2025, predating Anthropic's formalization of the hooks contract in Claude Code 1.0.51. Claude-flow's hooks vocabulary is a public reference for how agent-tool lifecycle can be modeled. Strategic significance: Early codification of agent-tool lifecycle as a first-class contract. The substrate everything else hooks into. First-class agent memory exposed as memory store / memory search at the CLI, backed by AgentDB with HNSW ANN search. Measured 1.9x at N=20k, 3.2-4.7x at N=5k vs brute force, recall@10 ~0.99. First commit: — 2025-10-20 2caa31c1a AgentDB v1.3.9: — 2025-10-23 173f818bb 8 controllers + 6 MCP tools: — 2026-02-26 3a63e1ab5 npm: claude-flow@2.7.0-alpha.14 2025-10-20 First-claim status: First in the Claude-Code harness category for HNSW-indexed agent memory exposed via memory store / memory search at the CLI surface. Vector DBs predate this; embedding one as the agent's primary memory plane with measured benchmark evidence in October 2025 is the harness-level first. Strategic significance: Most agent frameworks store memory as opaque JSON or rely on a separate RAG service. HNSW-indexed agent memory at the CLI was unusual in October 2025. 32x memory reduction with reconstruction-cosine ≥ 0.99999 — measured, not claimed. Shipped in a developer-facing agent CLI, not just a vector DB. First commit: — 2026-04-28 ca4d1f0a4 Guidance plane 2.7x win at N=1000: — 2026-05-22 cff24dbff npm: ruflo@3.6.5 2026-04-28 , claude-flow@3.6.5 same day First-claim status: First in the Claude-Code harness category for embedded 1-bit quantization with measured reconstruction-cosine. Research-grade quantization RaBitQ, PQ predates this; embedding it directly in a developer-facing agent CLI with measured benchmark evidence at compile time is the harness first. Strategic significance: Putting 1-bit quantization directly in the agentic loop — not in a downstream DB — collapses memory footprint without leaving the orchestrator. A pattern store with explicit JUDGE/DISTILL phases — Q-learning-adjacent — layered onto the core memory system. WASM-compiled for embedded use. First commit: — 2025-10-12 c31098e8a WASM integration: — 2025-10-13 b66d2daf9 V3 guidance system: — 2026-01-05 dac564a3e npm: claude-flow@2.7.0-alpha.6 2025-10-13 First-claim status: Among the earliest agent-CLI implementations of explicit trajectory-learning with JUDGE/DISTILL/CONSOLIDATE phases. Reinforcement-learning research predates this; productionizing a four-phase pipeline with WASM compilation inside an agent-orchestration CLI in October 2025 is, at compile time, ahead of comparable harnesses. Strategic significance: Ahead of similar features in competing harnesses. The pipeline names a clear feedback loop, which is what learning ultimately requires. Fast-apply merge engine for LLM-produced edit snippets, wired in front of Haiku/Sonnet/Opus rather than after. Edits at substrate speed, model only when needed. First commit: — 2025-10-12 bcf58260f Supporting npm: agent-booster@0.1.0 2025-10-08 npm in claude-flow: claude-flow@2.5.0-alpha.140 2025-10-09 First-claim status: First in the Claude-Code harness category for fast-apply merge wired as a pre-LLM tier with measured 352x speedup. IDE-assistant speed-edit caches predate this but apply post-LLM; positioning the fast-apply engine ahead of the model call is the harness first. Strategic significance: Fast-apply in front of the LLM not after is a different design from speed-edit caches popular in IDE assistants. It opened the deterministic-tier door that ADR-143 later walked through. Route a task to 1 deterministic WASM/codemod, 2 Haiku, or 3 Sonnet/Opus by complexity. Markers CODEMOD AVAILABLE and TASK MODEL RECOMMENDATION injected into the agent prompt. ADR added: — 2026-01-14 1636f2ee9 npm: claude-flow@3.0.0-alpha.x January 2026 First-claim status: First in the Claude-Code harness category to codify a three-tier routing policy with a $0 deterministic tier as Tier 1 — and to inject routing markers CODEMOD AVAILABLE , TASK MODEL RECOMMENDATION into the agent prompt as a programmatic signal. Model-routing libraries exist; specifying $0 Tier-1 in an architectural decision record with verifiable cost evidence is the harness first. Strategic significance: Early codification of "don't call the LLM if you don't have to" — Tier 1 at $0 quantifies the savings instead of asserting them. Plugin discovery, install, and updates served from an IPFS CID via Pinata gateway, with content-addressed publishing and an auto-fallback demo registry. First commit: — 2026-01-24 d2cf141c3 Live registry pushed: — 2026-01-24 963595e75 npm: claude-flow@3.0.0-alpha.171 2026-01-24 First-claim status: Verified first for an IPFS-anchored plugin registry serving an agent-orchestration CLI. IPFS-distributed package registries exist in adjacent domains; using one to distribute agent plugins with auto-fallback to a demo registry is, at compile time, unique to this project. Strategic significance: Most CLI plugin systems in early 2026 were centralized. IPFS-anchored content-addressed publishing gives the marketplace decentralized durability — no operator to seize. Real-time scanning of agent I/O for manipulation attempts AIMDS pattern , with 25-level meta-learning adaptive mitigation and OWASP ASI01 Indirect Prompt Injection mapping. First commit: — 2026-01-12 14c254763 ToolOutputGuardrail + OWASP 2026 ADR-131/132 : — 2026-05-26 0f5b8fe7f npm: claude-flow@3.0.0-alpha.x January 2026 First-claim status: First in the Claude-Code harness category for OWASP ASI01 Indirect Prompt Injection defense mapping with 25-level adaptive mitigation refined by a nightly research swarm. Prompt-injection scanners predate this; an agent CLI that researches its own threat model overnight and ships ADRs against it is, at compile time, unique. Strategic significance: Defenses mapped to an industry threat-model taxonomy — and refined by a nightly research swarm — make security claims auditable instead of marketing. A 5-phase development methodology with a CLI mode per phase and an orchestrator that sequences them. Cited in external coverage as "the thing that set claude-flow apart from other 'run multiple agents' tools." First commit: — 2025-06-11 1c09cd277 Slash-command integration: — 2025-06-13 95165e822 npm: claude-flow@1.0.32 2025-06-11 First-claim status: Verified first in the Claude-Code harness category, externally corroborated. The DEV Community write-up explicitly cites SPARC as the differentiator that set claude-flow apart from contemporary "run multiple agents" tools — an independent attribution of first-mover status to the SPARC methodology in agent CLIs. Strategic significance: Methodology, not feature. SPARC is what gave external observers a word for what claude-flow was actually doing. A claude-flow-codex orchestrator that spawns real codex exec workers alongside claude -p workers in a shared memory namespace — pre-built templates for feature, security, refactor, and bugfix work. First commit real wiring : — 2026-05-11 1b632ce8d First stable npm: claude-flow@3.7.0 2026-05-24 First-claim status: Verified first for cross-vendor host-harness orchestration spawning real codex exec workers alongside claude -p workers in a shared memory namespace. Multi-LLM routing libraries predate this; orchestrating two different vendor CLIs in coordinated work is, at compile time, unique. Strategic significance: Cross-vendor host orchestration. Most multi-platform frameworks abstract the LLM but not the host harness; this abstracts both. Cognitum Seed ESP32 devices as swarm members with 5-tier trust scoring; firmware rollouts via anomaly-gated canary advancement. First commit: — 2026-04-29 36e9b4d61 Real device bridge 24 tests : — 2026-04-29 ef7369aaf npm: @claude-flow/plugin-iot-cognitum@1.0.0-alpha.4 April 2026 First-claim status: Verified first for treating an IoT device as a swarm-member peer to LLM agents, with tier-scored trust and anomaly-gated firmware rollouts. IoT device-management platforms exist; embedding a physical device as a co-equal agent in a multi-agent harness is, at compile time, unique. Strategic significance: Cross-domain swarm membership — an ESP32 in the same swarm as an LLM agent — is novel. Most agent frameworks have no physical-device primitive. Multi-agent payment authorization with cryptographic verification and Byzantine consensus. Payment authorization with BFT inside an agent CLI is unusual; most "agent commerce" was prototype-level in late 2025. First commit: — 2025-09-30 8473e8b5a Supporting npm: agentic-payments@1.0.0 2025-09-30 npm in claude-flow: @2.5.0-alpha.x series September 2025 First-claim status: Among the earliest payment-authorization primitives with BFT consensus inside an agent CLI. Agent-commerce demos appeared in late 2025; shipping it as an MCP integration with cryptographic verification predates the broader agent-commerce wave. Strategic significance: Real commerce primitives — not demos — inside an agent harness. Per-session, per-model, per-agent USD attribution with a corpus benchmark that produces JSON evidence. Federation-wide budget circuit breaker ADR-097 extends the same accounting to peers. First commit: — 2026-04-30 f25bcf207 Agent Booster verified: — 2026-05-05 0fd61e3e5 npm: @claude-flow/plugin-cost-tracker April 2026 First-claim status: First in the Claude-Code harness category for verifiable JSON-evidence cost attribution per session/model/agent — and the only agent CLI at compile time with a federation-wide budget circuit breaker ADR-097 that propagates spend limits across peer installs. Strategic significance: Cost reports tied to a verifiable benchmark JSON — rather than provider invoices — is rare in agent CLIs. Bee-themed multi-worker coordination with a queen-led leader, persistent SQLite state, and resume capability. 75% performance improvement landed within hours of the initial Hive-Mind merge. First commit: — 2025-07-06 a1f85645b Resume capability: — 2025-07-14 af1c8a7c7 npm: claude-flow@2.0.0-alpha.2 2025-07-07 First-claim status: Among the earliest queen-led multi-agent harnesses with persistent SQLite state and resume across sessions. Cited in Issue 113 https://github.com/ruvnet/claude-flow/issues/113 as "Claude Flow v2.0.0 Revolutionary AI Swarm Orchestration Platform" — independent contemporaneous framing of the v2.0 release as category-defining. Strategic significance: Persistent, resumable multi-agent sessions with leader election were rare in agentic CLIs in July 2025. A scheduled swarm whose output is a commit. Surfaces gaps e.g. the May 26 commit that flagged a 5-decision gap in RL orchestration across HAL/AutoGen/etc. before they become incidents. First commit: — 2026-05-26 3de265e76 Follow-up: — 2026-05-27 b3b96010e npm: claude-flow@3.10.x late May 2026 First-claim status: Verified first for a self-auditing agent harness whose nightly research output is a versioned, dated commit on its own repository. Continuous-improvement pipelines exist in MLOps; turning an agent CLI's gap-analysis into a public git artifact, on a schedule, is, at compile time, unique. Strategic significance: Treating the framework's own gap-analysis as a versionable, dated, signed artifact is unusual practice. The harness audits itself, on a schedule, and the audit is a commit. The trajectory is the story. Each quarter is a layer; each layer is the foundation for the next. Q2 2025 ORCHESTRATION │ Repo initialised Jun 02 │ First orchestration system + MCP Jun 10, claude-flow@1.0.0 │ Swarm subsystem Jun 11, @1.0.22 │ SPARC methodology Jun 11, @1.0.32 │ MCP server + SPARC integration Jun 19, @1.0.72 ▼ Q3 2025 HOOKS + CONSENSUS │ ruv-swarm MCP coordination Jul 01 │ Hive-Mind v2.0.0 queen-led Jul 06, @2.0.0-alpha.2 │ Complete hooks system Jul 08, @2.0.0-alpha.16 │ Hook safety system Jul 08, @2.0.0-alpha.27 │ 64-agent + consensus suite Jul 25, @2.0.0-alpha.73 │ Maestro multi-agent conductor Jul 29 │ agentic-payments BFT Sep 30, @2.5.0-alpha.x ▼ Q4 2025 MEMORY + ROUTING │ PreToolUse modification hooks Oct 09, @2.5.0-alpha.140 │ Agent Booster 352x Oct 12 │ ReasoningBank optional memory mode Oct 12 │ ReasoningBank WASM Oct 13, @2.7.0-alpha.6 │ AgentDB skills + HNSW Oct 20, @2.7.0-alpha.14 │ AgentDB v1.3.9 + 3-agent swarm Oct 23 │ MCP 2025-11 spec compliance Nov 12, @2.7.34 ▼ Q1 2026 V3 — UNIFIED PLATFORM │ Deep agentic-flow integration roadmap Jan 03 │ V3 ReasoningBank guidance + CLI Jan 05 │ First @claude-flow/cli publish Jan 06, @3.0.0-alpha.1 │ First claude-flow v3 publish Jan 06, @3.0.0-alpha.1 │ Workers system, ADR-014 Jan 07 │ Transfer-hook IPFS pattern sharing, ADR-021 Jan 08 │ AIDefence, ADR-022 Jan 12 │ 3-tier model routing, ADR-026 Jan 14 │ SONA WASM + SemanticRouter Jan 24, @3.0.0-alpha.170 │ Live IPFS plugin registry Jan 24, @3.0.0-alpha.171 │ Statusline observability, ADR-052 Feb 10 │ First ruflo wrapper publish Feb 16, ruflo@3.1.0-alpha.14 │ AgentDB 3.0.0-alpha.9 + 8 controllers Feb 26 │ RVF cognitive container format Feb 28 │ RuVector WASM + semantic embeddings Mar 17 ▼ Q2 2026 FEDERATION + VERIFICATION │ RVM execution substrate, ADR-074 Apr 06 │ Claude Code memory bridge, ADR-076 Apr 06 │ 19 native plugins, browser harness Apr 27 │ ruflo-goals GOAP + horizon tracking Apr 27 │ RaBitQ 1-bit quantization Apr 28, ruflo@3.6.5 │ Agent LLM Federation, ADR-078 Apr 29 │ IoT Cognitum, ADR-079 Apr 29 │ 10 new ruflo plugins Apr 30 │ Federation budget circuit breaker, ADR-097 May 03 │ Witness manifest CI verification May 08 │ ADR-103 witness temporal history May 08 │ Real dual-mode Claude+Codex May 11 │ Signed browser trajectories, ADR-122 May 18 │ First v3 stable release May 24, claude-flow@3.7.0 │ ToolOutputGuardrail + OWASP 2026 May 26 │ Nightly dream swarm research May 26 │ MoE expert routing, ADR-135 Track G May 27 │ Deterministic Tier-1 codemods, ADR-143 May 29 ▼ Today v3.10.46 Jun 13, 2026 The evolution narrative reads, in five words: orchestration → hooks → memory → routing → federation/verification. Each layer is permission for the next. You cannot federate trust without provenance; you cannot route without observability; you cannot remember without a memory plane. Year one built the stack bottom-up. What plugs into claude-flow, and what claude-flow plugs into. The harness is a hub, not a monolith. | Layer | Component | Role | |---|---|---| | Host harness | Claude Code, Codex | Interactive + headless execution surfaces | | Protocol | MCP Model Context Protocol 2025-11 spec | Tool catalog contract | | Models | Anthropic Claude Opus, Sonnet, Haiku | Tier 3 / Tier 2 routing destinations | | Models | OpenAI Codex via codex exec | Dual-mode worker pool | | Models | agentic-flow @alpha ONNX backend | Embeddings all-MiniLM-L6-v2, 384-dim | | Domain | Components | |---|---| | Memory plane | AgentDB 19 controllers , RuVector HNSW + RaBitQ , sql.js SQLite cache, ReasoningBank trajectory store, RVF cognitive containers | | Intelligence pipeline | SONA 0.0043ms/adaptation measured , MoE 8 experts, convergence measured , Flash Attention target only — unverified , EWC++ catastrophic-forgetting prevention | | Hooks | 17 hook types pre/post-edit, pre/post-task, session-, route, explain, transfer, intelligence-trajectory- | | Workers | 12 background workers ultralearn, optimize, consolidate, predict, audit, map, preload, deepdive, document, refactor, benchmark, testgaps | | Security | AIDefence AIMDS pattern, 25-level meta-learning , ToolOutputGuardrail, witness manifests Ed25519 , OWASP ASI01 mapping | | Distribution | IPFS via Pinata plugin registry , GitHub source + releases , npm 3 packages × 3 dist-tags | | Category | Plugins | |---|---| Architecture & governance | ruflo-adr ADR lifecycle , ruflo-ddd domain-driven design , ruflo-sparc 5-phase methodology , ruflo-cost-tracker USD attribution + budgets , ruflo-observability traces, metrics, logs | Memory & intelligence | ruflo-agentdb, ruflo-rag-memory, ruflo-ruvector, ruflo-knowledge-graph, ruflo-intelligence, ruflo-ruvllm | Coordination | ruflo-swarm, ruflo-federation, ruflo-loop-workers, ruflo-goals GOAP + horizon tracking | Safety & verification | ruflo-aidefence, ruflo-security-audit, ruflo-jujutsu signed git workflows | Cross-domain | ruflo-browser RVF-signed sessions , ruflo-iot-cognitum device-as-agent , ruflo-neural-trader LSTM/Transformer + Rust backtest , ruflo-market-data, ruflo-rvf cognitive containers | | Channel | Surface | |---|---| npm latest | npx claude-flow@latest , npx @claude-flow/cli@latest , npx ruflo@latest | npm alpha / v3alpha | Legacy compatibility — same version as latest | IPFS Pinata | Plugin registry, content-addressed JSON manifests | GitHub releases | gh release create v